feat(minimax-h3): DiffusionNFT on the vLLM-Omni rollout backend at 32-GPU scale - #5

Draft

nussejzz wants to merge 39 commits into

feat/minimax-h3-vllmomni-32gpufrom

feat/h3-diffusionnft-vllmomni

Commits on Aug 26, 2026

feat(minimax-h3): t2va model package for Flow-GRPO (trainside)

celve
authored and
nussejzz
committed
e303f0a
fix(minimax-h3): make the bundle a Remote and stop reading .type off bundle.device

celve
authored and
nussejzz
committed
5ba438f
fix(minimax-h3): pipeline must be a Remote; derive replay geometry from params

celve
authored and
nussejzz
committed
f8d1cf4
fix(minimax-h3): root_wrap is an FSDPConfig field, not an FSDPBackend one

celve
authored and
nussejzz
committed
0261153
fix(minimax-h3): scope LoRA to transformer_blocks so no adapter lands outside FSDP

celve
authored and
nussejzz
committed
48ace8d
fix(minimax-h3): locate Qwen3-VL decoder layers and run the encoder on its own device

celve
authored and
nussejzz
committed
25f138c
fix(minimax-h3): read prompts off Texts.texts, not Texts.to_list()

celve
authored and
nussejzz
committed
29d3af3
fix(minimax-h3): audio latent width, packed batch axis, decode channels, trajectory positions

celve
authored and
nussejzz
committed
7d0ed4f
fix(minimax-h3): build denoise generators per step with the real signature

celve
authored and
nussejzz
committed
3fd2dad
perf(minimax-h3): keep the VAEs on the train device, park only the conditioner

celve
authored and
nussejzz
committed
82e56b2
fix(minimax-h3): sample_ids live on the Part, not the Sample

celve
authored and
nussejzz
committed
e1d18a9
fix(minimax-h3): generate() must return a Sample, not the filled Part

celve
authored and
nussejzz
committed
a073770
fix(reward/clap): ClapProcessor takes `audio`, not the removed `audios`

celve
authored and
nussejzz
committed
4f46ce6
fix(minimax-h3): replay a subset of timesteps -- all 10 does not fit on 95 GB

celve
authored and
nussejzz
committed
5b71cf1
fix(minimax-h3): eta 0.7 destroys this checkpoint -- drop to 0.1

celve
authored and
nussejzz
committed
a217f61
tune(minimax-h3): eta 0.1 -> 0.2, chosen on measured reward spread not eyeballing

celve
authored and
nussejzz
committed
64e3f40
feat(reward): let videopickscore score a mid-clip frame instead of frame 0

celve
authored and
nussejzz
committed
5098f70
tune(minimax-h3): group size 2 makes every advantage exactly +-1 -- default to 8

celve
authored and
nussejzz
committed
9522da8
style(reward): ruff-format the t2av override comprehension

celve
authored and
nussejzz
committed
e4a9281
feat(minimax-h3): switch to CPS at eta 0.7 -- 4.2x the exploration at better quality

celve
authored and
nussejzz
committed
7ca6c4f
fix(minimax-h3): adam_epsilon 1e-8 exceeds the per-coordinate gradient -- drop to 1e-12

celve
authored and
nussejzz
committed
61b8c7a
fix(reward/videoalign): AutoModelForTextToVideo was removed in transformers>=5

celve
authored and
nussejzz
committed
49884b9
fix(reward/videoalign): Qwen2VLConfig nests hidden_size under text_config in transformers>=5

celve
authored and
nussejzz
committed
b0fd7b4
fix(reward/videoalign): checkpoint load was a silent no-op -> constant 0.0 rewards

celve
authored and
nussejzz
committed
2a231c7
feat(minimax-h3): add vLLM-Omni rollout at 32-GPU scale

nussejzz
and
cursoragent
committed
72e7732
chore(vllm-omni): migrate H3 rollout to current runtime

nussejzz
and
cursoragent
committed
3e87f52
fix(deps): keep engine-specific Transformers bounds

nussejzz
and
cursoragent
committed
4de8661
chore(deps): follow rebased vLLM-Omni integration

nussejzz
and
cursoragent
committed
1a2c32b
style(minimax-h3): satisfy current repository policy

nussejzz
and
cursoragent
committed
4e9e7ae
docs(minimax-h3): preserve model contracts after lint migration

nussejzz
and
cursoragent
committed
7de06e6
fix(vllm-omni): align H3 with the current CUDA stack

nussejzz
and
cursoragent
committed
669baa5
feat(weight-sync): verify H3 LoRA actually landed in the fused rollout layout

nussejzz
and
cursoragent
committed
bc8b59f
fix(minimax-h3): make the vLLM-Omni rollout contract exact

nussejzz
and
cursoragent
committed
90e5fae
feat(reward): score video over time and fail closed on partial T2AV rewards

nussejzz
and
cursoragent
committed
e49c1d6
feat(minimax-h3): 32-GPU hybrid-parallel recipes and continuation guarantees

nussejzz
and
cursoragent
committed
12e9c1a
fix(train): make EMA LoRA training work with a colocated external rollout

nussejzz
and
cursoragent
committed
e4d3c72
feat(minimax-h3): forward-process stage API for DiffusionNFT

nussejzz
and
cursoragent
committed
99d5751
feat(minimax-h3): record a forward-process rollout on the vLLM-Omni backend

nussejzz
and
cursoragent
committed
da57629
feat(minimax-h3): 32-GPU DiffusionNFT recipe on the vLLM-Omni backend

nussejzz
and
cursoragent
committed
4aca1ef